Assets

An asset is one entry in the assets map, keyed by a reference string and pointing at an image sent either as a multipart file part or as base64 in the JSON body. The reference string is then placed in state, instructions, or criteria, and the API performs a find-and-replace, swapping each reference for the image before the model reads it.

Assets are experimental and gated: every request containing an image or a Box question needs the TypeSafe-Danger-Zone-Images header, images are PNG or JPEG stills from 256 to 1568 pixels per side, and each use of a reference counts toward the limits again. The docs recommend keeping the simplest possible question shape, because all models perform worse on images than on text.